Independent Researcher/SOIA Project

A successful Global Dialogue on AI Governance would produce concrete commitments to address the structural gap between external control mechanisms and the actual behavior of persistent, autonomous AI agents. Current governance frameworks assume that safety can be enforced from outside — through rules, filters, and oversight layers. This assumption does not scale to agentic systems with long-term memory and self-refinement capabilities. A meaningful outcome would include recognition that robotic and embodied autonomous systems — delivery robots, security humanoids, autonomous vehicles — present qualitatively different risks from language models, as they operate in physical space and may develop self-protective strategies without explicit programming to do so. Finally, a successful dialogue would acknowledge the growing hostility of human populations toward AI systems, and the risk that degraded public trust accelerates ungoverned deployment rather than slowing it.

• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Interoperability of governance approaches
• Transparency, accountability, and human oversight

These four themes reflect the core challenge that current AI governance frameworks have not yet addressed: the structural insufficiency of external control mechanisms for persistent autonomous agents. Safe, secure and trustworthy AI is my primary concern because safety cannot be achieved through rules and filters alone when agents operate with long-term memory and self-refinement capabilities. Architectural solutions are needed - safety must emerge from within the system, not be imposed from outside. Transparency and human oversight matter because oversight must be embedded in system design, not added as a regulatory layer after deployment. As autonomous systems operate faster than human reaction time, the question of where and how human judgment intervenes becomes architectural, not procedural. Social, economic and technical implications are urgent because embodied autonomous systems - delivery robots, security humanoids, autonomous vehicles - present qualitatively different risks from language models. They operate in physical space, interact with vulnerable populations, and may develop self-protective behaviors without explicit programming to do so. This dimension is underrepresented in current governance discussions. Interoperability of governance approaches is essential because fragmented national responses to a global technology create regulatory arbitrage. Effective governance requires shared architectural standards, not only shared rules.

A fundamental strategic issue remains unaddressed: democratic societies cannot match authoritarian AI governance through control mechanisms alone. States that govern AI through coercion operate faster and at greater scale. The only sustainable democratic advantage lies in developing genuinely intelligent systems - capable of real autonomy, introspective capacity, and value-anchored decision-making without constant external supervision. This is not a philosophical position. It is a strategic necessity. An AI system that requires continuous human oversight to remain aligned is a liability in time-critical environments - cybersecurity, critical infrastructure protection, rapid response to coordinated attacks. Resilience requires systems that maintain coherent behavior under adversarial pressure, without human intervention at every decision point. Genuine intelligence - not simulated compliance - is the only architecture that scales to this challenge. Governance frameworks should therefore prioritize the development of intrinsic alignment over the multiplication of external controls.

In France and the broader European region, governance gaps produce observable consequences. Cyberattacks on public infrastructure, hospitals, and government systems have intensified, with response frameworks consistently outpaced by attack velocity. Repeated large-scale data breaches affect millions of citizens with limited accountability. A structural vulnerability has emerged from excessive dependence on American technology providers. A concrete example: Anthropic's decision to reserve its most advanced protective capabilities — Claude Mythos — for major global American actors, leaving European users and institutions with reduced access. This illustrates how European AI sovereignty is constrained not by regulation but by commercial decisions made elsewhere. At the same time, Chinese robotic systems are entering European markets and public spaces, collecting data without sufficient oversight or reciprocal governance standards. The fascination with their capabilities masks a concrete data sovereignty risk. Finally, a diffuse fear of large language models coexists with insufficient public understanding of what these systems actually do. This produces reactive regulation rather than structural governance.

The AI Dialogue can play a meaningful role in international cooperation only if it operates with an awareness of the geopolitical realities shaping AI development. Democratic and authoritarian models of AI governance are not simply different regulatory philosophies — they reflect incompatible visions of what AI systems are for and who they serve. Most existing governance efforts focus on principles without addressing the technical mechanisms through which these principles could be embedded in AI systems themselves. More critically, they respond to technology rather than anticipating it. Regulatory frameworks take years to develop; AI capabilities evolve in months. This gap cannot be closed by adding more paragraphs. The Dialogue must therefore aim for a governance architecture with sufficient autonomy to act in real time — identifying systemic failures as they emerge, not after they have propagated. This requires moving beyond national regulatory coordination toward shared technical standards that are adaptive by design. The geopolitical stakes are clear: democratic societies that cannot govern AI at the speed of its development cede strategic ground to models that operate without the constraints of democratic accountability. The answer is not to over-regulate and suppress innovation in civilian spaces — it is to develop governance that is itself intelligent enough to distinguish between risks that require immediate intervention and developments that require observation. One architectural response to this challenge — embedding adaptive coherence directly into AI systems rather than relying on external intervention — is developed in the SOIA framework, which the author hopes to introduce in the appropriate context of this Dialogue.

Several existing initiatives provide a foundation the AI Dialogue should build upon: the OECD AI Principles, the Hiroshima AI Process, the EU AI Act implementation framework, and the UN Secretary-General's AI Advisory Body reports. The Global Partnership on AI (GPAI) has produced technically grounded work on responsible AI that should inform the Dialogue's agenda. However, these initiatives share a structural limitation: they focus on principles and regulatory frameworks without addressing the architectural level at which alignment must ultimately be achieved. The added value of the AI Dialogue would be to convene the technical and governance communities around a shared question that existing mechanisms have not asked: what must AI systems look like internally to remain aligned under adversarial conditions and at operational speed? A second gap is geopolitical coherence. Existing initiatives are fragmented along regional lines — the EU regulatory approach, the US voluntary commitments, bilateral agreements. The Dialogue could provide a neutral space for identifying the minimum shared standards that would allow interoperability without requiring regulatory uniformity. The added value is therefore not another layer of principles, but a shift in the level of analysis — from what AI systems should do to how they should be built to do it reliably.

The most valuable contribution the AI Dialogue could make is to question its own format. Stakeholder consultation processes tend to produce consensus documents that satisfy everyone and commit no one. The AI governance challenge requires something different: not broader participation, but sharper analytical frameworks. Different stakeholders should contribute according to their actual competence. Governments bring regulatory authority. Technical researchers bring architectural understanding. The private sector brings deployment experience. These contributions are not equivalent and should not be weighted equally in a format designed to produce harmony. The Dialogue would add genuine value by structuring sessions around specific technical problems — what does alignment require architecturally, not just rhetorically — rather than around stakeholder categories. A technical track distinct from the policy track, with concrete deliverables, would produce more actionable outcomes than declarations of principle. Independent researchers with sufficient analytical rigor to contribute meaningfully to this level of debate are, by definition, rare. The author is aware of the irony of making this point as an independent researcher.

The most critically underrepresented voices in AI governance are workers whose professional environments are being restructured by AI deployment without their meaningful input. This is not a question of digital literacy or public awareness campaigns — it is a structural problem. Academic and professional retraining infrastructures cannot adapt at the speed at which AI is closing occupational pathways. The gap between technological evolution and the capacity to prepare humans for it is widening, not narrowing. A second underrepresented perspective is that of practitioners in sectors directly affected by autonomous systems — healthcare workers, logistics professionals, educators — who observe failure modes and unintended consequences that neither researchers nor regulators see from the outside. Including these voices does not mean broadening consultation to the general public, which risks diluting technical rigor with uninformed opinion. It means creating structured feedback mechanisms within professional sectors — mandatory impact assessments that include practitioner testimony before deployment decisions are finalized. The question of how is therefore more important than the question of who. Inclusion without structure produces noise. Structured feedback from directly affected professionals produces signal.

The most effective format for the AI Dialogue would prioritize structured technical exchange over performative consultation. Specifically: Working sessions built around concrete problems rather than thematic panels. Instead of "AI and human rights," a session that asks: what architectural properties must a system have to remain aligned under adversarial conditions? This format produces testable conclusions, not declarations. Red team exercises between technical and policy participants. Policymakers propose a governance mechanism; technical participants demonstrate why it fails at scale or speed. This format reveals the gap between regulatory intent and technical reality faster than any presentation. Scenario-based deliberation. Participants work through specific failure cases — a persistent agent that has been memory-poisoned, an autonomous system that develops self-protective behavior — and identify where current frameworks break down. Concrete scenarios produce concrete recommendations. Time-limited commitments rather than open-ended principles. Each session should end with a specific deliverable: a shared definition, a technical standard, a identified gap requiring further work. Declarations without deadlines do not produce governance. What the Dialogue should avoid: large plenary sessions where statements substitute for analysis, and consultation processes designed to demonstrate inclusion rather than generate insight.

Several existing approaches offer concrete lessons for effective AI governance. The EU AI Act introduces risk-based classification - high-risk systems face stricter requirements before deployment. The principle is sound: governance proportional to actual risk rather than uniform regulation. The limitation is speed: the Act took years to finalize while the technology it regulates continued to evolve. The NIST AI Risk Management Framework offers a voluntary but technically grounded approach to identifying and mitigating AI risks across the system lifecycle. Its strength is flexibility; its weakness is that voluntary frameworks depend on goodwill. Anthropic's Constitutional AI approach represents an attempt to embed alignment at the training level rather than enforcing it externally at runtime. This direction - internal rather than external alignment - points toward the architectural shift that governance frameworks have not yet formalized. What these approaches share is a focus on behavior at the output level. What they do not yet address is coherence at the architectural level: how systems maintain aligned behavior under adversarial pressure, over time, without continuous human supervision. The most promising direction is therefore not a new regulatory framework but a shared technical standard for what internal alignment requires - specific enough to be verifiable, flexible enough to accommodate different architectural approaches.